Worked on the apple/axlearn repository to enhance reproducibility in machine learning experiments by introducing deterministic sampling to the top_k_logits function. Focused on data processing and Python development, the work involved adding a new parameter to control tie-breaking when k equals 1, ensuring that sampling outcomes could be reproduced reliably. The implementation updated the function signature and refined the tie-breaking logic to return either all tied logits or the smallest index, depending on the parameter. Comprehensive test coverage was added to validate deterministic behavior and edge cases, reflecting a methodical approach to improving model evaluation consistency using Python and machine learning techniques.